Light Weight Semantic Segmentation: A Modified DDRNET Approach Trained on Cityscapes and COCO-Stuff Datasets for Efficient Image Analysis

This paper presents research findings on a lightweight semantic segmentation model. The objective of this research was to train and test a semantic segmentation model on the Cityscapes and COCO-Stuff datasets, with the aim of achieving a minimum performance speed of 20 and 10 frames per second (FPS) on target CPU and mobile devices, respectively. Various models were implemented and evaluated for semantic segmentation, and among them, DDRNET-32-slim exhibited the best performance. Therefore, solutions were proposed and implemented to improve DDRNET-32-slim. The corresponding experiments demonstrated that the enhanced model achieved improved speeds of 28 and 15 FPS on CPU and mobile devices, respectively.